A unified soil thermal conductivity model based on artificial neural network

An accurate prediction of thermal conductivity (k) of unsaturated soils is a challenging problem in geothermal applications because k is affected by various factors such as soil fabric, water content, dry density, soil mineralogy and type. This study proposes a unified soil thermal conductivity model to consider these influence factors based on a screened artificial neural network (ANN) approach. Two hundred and fifty-seven (257) thermal conductivity measurements of unsaturated soils were first collected from the literature to compile a database. The ANN approach constrained with model complexity optimization and monotonicity control was then utilized to overcome the potential sample insufficiency problem in order to establish reliable correlations between k (represented by an empirical coefficient) and its influence factors including dry density (rho(d)), porosity (n), saturation degree (S-r), quartz content (m(q)), sand content (m(s)) and clay content (mc). Using this approach, some easy-to-use ANN-based contours for predicting k were developed to guide engineering practice. Results revealed that ANN-based predictions of k by the developed model matched experimental data with enhanced accuracy as compared with some benchmarking thermal conductivity models. The model can rationally consider the effects of multiple influence factors and their coupling effects on k in a quantitative and systematical way.